Is there any solution making nearest interpolation do not return zero grad?

autograd
1

Hello,

I am using optical flow network to calculate flows between frames, but when the nearest interpolation mode is adopted, it always return 0 grad back to optical flow network.

I know bilinear mode is commonly adopted. But in my case nearest sampling is necessary, so is there any solution making nearest interpolation mode do not return 0 grad?

Thanks in advance!

2

I see that nearest mode performs gradient backward correctly in the following example:

import torch, torch.nn.functionasl as F
x = torch.randn(1,1,5,5,requires_grad=True)
y = F.interpolate(x, size=(10,10), mode='nearest')
y.sum().backward()
print(x.grad)

Can you check if any other operation that you use breaks the computation graph or if the gradients are really 0 from the later layers?